Latent Space

Chai Discovery on turning drug discovery into an engineering discipline

Matt McPartlon & Neil Patil· Co-founder & product lead at Chai Discovery
·~95 min·English·Latent Space
AI CompanyTrainingReasoningOpen Source
TL;DR

Chai Discovery's research and product leads explain how folding and design models turned antibody discovery from blind trial-and-error into declarative precision engineering — and why they sell the models, not the drugs.

01Core Problem

Feeling Around in the Dark

<strong>Biology's real bottleneck isn't computing the answer — it's that you can't see the problem or cheaply check it, and structure prediction cracked the seeing half.</strong>

just how much of it is literally feeling around in the dark and that's not even a metaphor. You literally can't see like how these things look, right?

— Matt McPartlon, Latent Space
Key Insight
This reframes what 'AI for science' wins: not a cleverer search over a known space, but building the instrument that makes the space visible at all — the way the microscope had to precede microbiology.
02How the Models Work

Read, Then Write

<strong>Chai-1 predicts a structure from a sequence; Chai-2 is a different, all-atom diffusion model that co-designs a binder's sequence and 3D shape for a target you choose.</strong>

a nice thing with diffusion is like you can do this pretty slowly and pretty iteratively. So you can give the model a lot of time to think about all right if I change the structure like this how should the sequence change

— Matt McPartlon, Latent Space
Key Insight
Crossing from prediction to design is the same generative leap LLMs made — from labeling the world to producing it — except a molecule only works if its sequence and its 3D shape are mutually self-consistent, so both have to be generated together.
03The Proof

The 50-Target Bet

<strong>Chai pre-committed to designing antibodies against 50 externally-validated targets and got binders on about half — the statistic that flipped pharma from skeptical to interested.</strong>

we chose 50 targets designed antibodies against them. uh got hits to half and at that point I think pharma starts to realize like okay there actually signs of life here and this this might actually work in some of our programs

— Matt McPartlon, Latent Space
Key Insight
The discipline is in the setup, not the result: 50 targets picked from a contract lab's already-validated catalog and committed to in advance turns a cherry-picked demo into a hit rate a drug company can underwrite.
04The Product

Photoshop for Proteins

<strong>The product deliberately looks like Figma or Photoshop, not ChatGPT</strong> — because designing a molecule is a visual, spatial act, not a chat.

There's this almost like Photoshopesque like design suite. You have this equivalent of a paint tool to kind of paint your epitope. You have this equivalent of a contentaware fill tool to kind of get your uh your binders generated from Chai.

— Neil Patil, Latent Space
Key Insight
Choosing a spatial CAD interface over a chatbot is a bet about where the work actually is: the hard part isn't phrasing a request, it's steering a 3D binding decision — and a chat box hides exactly the geometry a scientist needs to see.
05Business Model

The Neutral Software Factory

<strong>Chai makes no drugs of its own — it's a neutral modeling layer, walled off per partner, whose incentives compound rather than compete: better models make partners win, which funds better models.</strong>

I love the incentive alignment between like you know we make the models better, the partners succeed more and just like you know that iterates on itself.

— Matt McPartlon, Latent Space
Key Insight
Refusing to run its own drug pipeline trades lottery-ticket upside for a compounding moat: Chai captures learning and trust across every partner's whole portfolio, and — like the enterprises that use Anthropic without handing over their data — it can improve without training on any single partner's secrets.
06Research Culture

Delete, Delete, Delete

<strong>Chai's research edge is aggressive simplicity: complexity and the bitter lesson are treated as fundamentally at odds.</strong>

One thing that I like to say is kind of like complexity and being bitter lesson pill they're like fundamentally at odds.

— Matt McPartlon, Latent Space
Key Insight
The nuance most people miss: the bitter lesson doesn't mean delete every inductive bias. AlphaFold's triangle layers are a bias worth their steep compute cost — the skill is knowing which few to keep and which to throw away, not scaling blindly.
07Infrastructure

The Compute Market Is LLM-Pilled

<strong>Modern GPUs and clusters were shaped for LLM workloads, and folding models — small hidden dimensions, huge sequence dimensions — are almost the opposite, leaving biology-specific performance on the table.</strong>

not only are they like costly in terms of compute, they're just like not efficient on modern GPUs either. You have small hidden dimensions, large sequence dimensions, like it's like exactly the opposite of what GPUs are designed to process.

— Matt McPartlon, Latent Space
Key Insight
Folding models — with their L-squared pair representations and tiny hidden dimensions — get pinned to the memory-bandwidth wall, where even a layer norm can stall, so hardware and kernels tuned to feed dense LLM matmuls starve the very ops these models lean on.
08The Worldview

From Experiment to Engineering

<strong>The whole thesis: biology is crossing the same threshold software and chip design already crossed — from blind trial-and-error to declarative precision engineering.</strong>

the mission of the company is to really turn you know drug discovery from a scientific experiment to an engineering discipline

— Neil Patil, Latent Space
Key Insight
When a field finally gets a CAD tool, its bottleneck stops being 'can we make it' and becomes 'what should we make' — which is exactly why Chai keeps reframing every scientist and engineer as a capital allocator deciding where to spend scarce attention and compute.